[XviD-devel] Hadamard Transform

Marco Al xvid-devel@xvid.org
Fri, 6 Sep 2002 17:41:34 +0200


Christoph Lampert wrote:

> That might be. But Hadamard made it into H.26L for mode decision

AFAICS they use it for subpixel refinement too! Which seems to be the only real
application where it would indeed give a pressing need for speedup.

I dont think the hadamard transform will be quite fast enough not to slow XviD
down significantly if it tried to use it for this though.

> , and
> since the encoder is free in which method to choose, I wanted to test it.
> _If_ the approximation is good and fast, other areas come to
> mind: Adaptive quantization

I dont see much advantage, you have to do the DCT transform anyway ... and you
can do anything in the DCT domain as easily as in the hadamard domain AFAICS.

It might be less work to calculate an image based distortion measure at a given
quantizer with the hadamard transform ... but calculating an approximation of
such measures in the DCT domain isnt hard either, while its harder to get an
exact image based error I think it will be a bit of a wash since the hadamar
transform in turn only simulates energy compaction of the DCT.
An approximation of the error should be good enough to compete.

> preprocessing

Dont quite see how.

> SKIP detection.

We dont know how to do it right in the DCT domain, let alone the hadamard one :)
Even so, the same arguement as for mode decision applies ... the DCT
computations themselves do not take significant time (unless XviD manages to be
sped up by over an order of magnitude).

Marco